Generative AI Methods for Wireless Propagation Prediction

The continuous expansion of wireless communication technologies creates a pressing need for intelligent planning of a plethora of existing and emerging wireless services, in a timely and cost-effective manner. Site-specific wireless propagation modeling, which is the prediction of signal levels generated by a wireless communication system in an environment, is an essential element of such an intelligent planning process. Propagation models can be derived by numerical algorithms such as ray-tracing or even full-wave methods for Maxwell’s equations, which are based on the physics of electromagnetic wave propagation. However, physics-based methods require significant computational resources. While trading accuracy for efficiency could have been an acceptable solution a few years ago, the emerging increasingly complex landscape of wireless services requires a proportional advancement in the state-of-the-art in propagation modeling: high-fidelity, simple to derive, computationally efficient models. Machine-Learning (ML) methods offer an alternative route to formulating propagation models that can potentially combine the efficiency of empirical with the accuracy of physics-based models, as long as they are “generalizable”. “Generalizable” neural network propagation models can “learn” the physics of radiowave propagation during training, thus becoming applicable to new problems with specifications (geometry, position, and type of transmitter/receiver antennas, and frequency of operation) that are beyond those included in the training set. Such “generalizable” models can clearly have a profound impact on propagation modeling studies, overcoming the classical dichotomy between computational efficiency and accuracy that has dominated this area. To this end, we discuss two types of generative AI models applied to indoor propagation scenarios at microwave frequencies: diffusion models and Generative Adversarial Networks (GANs). We introduce the basic features of each model, outlining similarities and differences to fully connected and Convolutional Neural Networks (CNNs). We show that diffusion models offer a combined solution to three important problems in propagation modeling: computing received signal strength maps in complex geometries in real-time; “generalizability”; effective integration of measured and synthetic data. Moreover, GANs are employed to refine the resolution of coarse received signal strength maps, significantly accelerating their computation. We compare these results to models based on U-nets (an efficient form of a CNN), illustrating the unique properties of generative models.